diff options
Diffstat (limited to 'erts/emulator/beam/erl_gc.c')
-rw-r--r-- | erts/emulator/beam/erl_gc.c | 277 |
1 files changed, 208 insertions, 69 deletions
diff --git a/erts/emulator/beam/erl_gc.c b/erts/emulator/beam/erl_gc.c index e2689f58c3..8ba94d89e9 100644 --- a/erts/emulator/beam/erl_gc.c +++ b/erts/emulator/beam/erl_gc.c @@ -1,7 +1,7 @@ /* * %CopyrightBegin% * - * Copyright Ericsson AB 2002-2012. All Rights Reserved. + * Copyright Ericsson AB 2002-2013. All Rights Reserved. * * The contents of this file are subject to the Erlang Public License, * Version 1.1, (the "License"); you may not use this file except in @@ -47,10 +47,6 @@ */ #define ALENGTH(a) (sizeof(a)/sizeof(a[0])) -static erts_smp_spinlock_t info_lck; -static Uint garbage_cols; /* no of garbage collections */ -static Uint reclaimed; /* no of words reclaimed in GCs */ - # define STACK_SZ_ON_HEAP(p) ((p)->hend - (p)->stop) # define OverRunCheck(P) \ if ((P)->stop < (P)->htop) { \ @@ -59,7 +55,7 @@ static Uint reclaimed; /* no of words reclaimed in GCs */ erts_fprintf(stderr, "htop=%p\n", (p)->htop); \ erts_fprintf(stderr, "heap=%p\n", (p)->heap); \ erl_exit(ERTS_ABORT_EXIT, "%s, line %d: %T: Overrun stack and heap\n", \ - __FILE__,__LINE__,(P)->id); \ + __FILE__,__LINE__,(P)->common.id); \ } #ifdef DEBUG @@ -120,6 +116,8 @@ static void offset_rootset(Process *p, Sint offs, char* area, Uint area_size, static void offset_off_heap(Process* p, Sint offs, char* area, Uint area_size); static void offset_mqueue(Process *p, Sint offs, char* area, Uint area_size); +static void init_gc_info(ErtsGCInfo *gcip); + #ifdef HARDDEBUG static void disallow_heap_frag_ref_in_heap(Process* p); static void disallow_heap_frag_ref_in_old_heap(Process* p); @@ -129,7 +127,7 @@ static void disallow_heap_frag_ref(Process* p, Eterm* n_htop, Eterm* objv, int n #if defined(ARCH_64) && !HALFWORD_HEAP # define MAX_HEAP_SIZES 154 #else -# define MAX_HEAP_SIZES 55 +# define MAX_HEAP_SIZES 59 #endif static Sint heap_sizes[MAX_HEAP_SIZES]; /* Suitable heap sizes. */ @@ -137,13 +135,42 @@ static int num_heap_sizes; /* Number of heap sizes. */ Uint erts_test_long_gc_sleep; /* Only used for testing... */ +typedef struct { + Process *proc; + Eterm ref; + Eterm ref_heap[REF_THING_SIZE]; + Uint req_sched; + erts_smp_atomic32_t refc; +} ErtsGCInfoReq; + +#if !HALFWORD_HEAP +ERTS_SCHED_PREF_QUICK_ALLOC_IMPL(gcireq, + ErtsGCInfoReq, + 5, + ERTS_ALC_T_GC_INFO_REQ) +#else +static ERTS_INLINE ErtsGCInfoReq * +gcireq_alloc(void) +{ + return erts_alloc(ERTS_ALC_T_GC_INFO_REQ, + sizeof(ErtsGCInfoReq)); +} + +static ERTS_INLINE void +gcireq_free(ErtsGCInfoReq *ptr) +{ + erts_free(ERTS_ALC_T_GC_INFO_REQ, ptr); +} +#endif + /* * Initialize GC global data. */ void erts_init_gc(void) { - int i = 0; + int i = 0, ix; + Sint max_heap_size = 0; ASSERT(offsetof(ProcBin,thing_word) == offsetof(struct erl_off_heap_header,thing_word)); ASSERT(offsetof(ProcBin,thing_word) == offsetof(ErlFunThing,thing_word)); @@ -155,9 +182,6 @@ erts_init_gc(void) ASSERT(offsetof(ProcBin,next) == offsetof(ErlFunThing,next)); ASSERT(offsetof(ProcBin,next) == offsetof(ExternalThing,next)); - erts_smp_spinlock_init(&info_lck, "gc_info"); - garbage_cols = 0; - reclaimed = 0; erts_test_long_gc_sleep = 0; /* @@ -168,22 +192,46 @@ erts_init_gc(void) * we really don't want that growth when the heaps are that big. */ - heap_sizes[0] = 34; - heap_sizes[1] = 55; - for (i = 2; i < 23; i++) { - heap_sizes[i] = heap_sizes[i-1] + heap_sizes[i-2]; + /* Growth stage 1 - Fibonacci + 1*/ + /* 12,38 will hit size 233, the old default */ + + heap_sizes[0] = 12; + heap_sizes[1] = 38; + + for(i = 2; i < 23; i++) { + /* one extra word for block header */ + heap_sizes[i] = heap_sizes[i-1] + heap_sizes[i-2] + 1; } + + /* for 32 bit we want max_heap_size to be MAX(32bit) / 4 [words] (and halfword) + * for 64 bit we want max_heap_size to be MAX(52bit) / 8 [words] + */ + + max_heap_size = sizeof(Eterm) < 8 ? (Sint)((~(Uint)0)/(sizeof(Eterm))) : + (Sint)(((Uint64)1 << 53)/sizeof(Eterm)); + + /* Growth stage 2 - 20% growth */ /* At 1.3 mega words heap, we start to slow down. */ for (i = 23; i < ALENGTH(heap_sizes); i++) { - heap_sizes[i] = 5*(heap_sizes[i-1]/4); - if (heap_sizes[i] < 0) { + heap_sizes[i] = heap_sizes[i-1] + heap_sizes[i-1]/5; + if ((heap_sizes[i] < 0) || heap_sizes[i] > max_heap_size) { /* Size turned negative. Discard this last size. */ i--; break; } } num_heap_sizes = i; + + for (ix = 0; ix < erts_no_schedulers; ix++) { + ErtsSchedulerData *esdp = ERTS_SCHEDULER_IX(ix); + init_gc_info(&esdp->gc_info); + } + +#if !HALFWORD_HEAP + init_gcireq_alloc(); +#endif + } /* @@ -216,7 +264,7 @@ erts_next_heap_size(Uint size, Uint offset) low = mid + 1; } } - erl_exit(1, "no next heap size found: %lu, offset %lu\n", (unsigned long)size, (unsigned long)offset); + erl_exit(1, "no next heap size found: %beu, offset %beu\n", size, offset); } return 0; } @@ -272,17 +320,6 @@ erts_heap_sizes(Process* p) return res; } -void -erts_gc_info(ErtsGCInfo *gcip) -{ - if (gcip) { - erts_smp_spin_lock(&info_lck); - gcip->garbage_collections = garbage_cols; - gcip->reclaimed = reclaimed; - erts_smp_spin_unlock(&info_lck); - } -} - void erts_offset_heap(Eterm* hp, Uint sz, Sint offs, Eterm* low, Eterm* high) { @@ -363,6 +400,7 @@ erts_garbage_collect(Process* p, int need, Eterm* objv, int nobj) Uint reclaimed_now = 0; int done = 0; Uint ms1, s1, us1; + ErtsSchedulerData *esdp = erts_get_scheduler_data(); #ifdef USE_VM_PROBES DTRACE_CHARBUF(pidbuf, DTRACE_TERM_BUF_SIZE); #endif @@ -370,11 +408,7 @@ erts_garbage_collect(Process* p, int need, Eterm* objv, int nobj) trace_gc(p, am_gc_start); } - erts_smp_proc_lock(p, ERTS_PROC_LOCK_STATUS); - p->gcstatus = p->status; - p->status = P_GARBING; - erts_smp_proc_unlock(p, ERTS_PROC_LOCK_STATUS); - + erts_smp_atomic32_read_bor_nob(&p->state, ERTS_PSFLG_GC); if (erts_system_monitor_long_gc != 0) { get_now(&ms1, &s1, &us1); } @@ -418,9 +452,8 @@ erts_garbage_collect(Process* p, int need, Eterm* objv, int nobj) ErtsGcQuickSanityCheck(p); - erts_smp_proc_lock(p, ERTS_PROC_LOCK_STATUS); - p->status = p->gcstatus; - erts_smp_proc_unlock(p, ERTS_PROC_LOCK_STATUS); + erts_smp_atomic32_read_band_nob(&p->state, ~ERTS_PSFLG_GC); + if (IS_TRACED_FL(p, F_TRACE_GC)) { trace_gc(p, am_gc_end); } @@ -445,11 +478,9 @@ erts_garbage_collect(Process* p, int need, Eterm* objv, int nobj) monitor_large_heap(p); } - erts_smp_spin_lock(&info_lck); - garbage_cols++; - reclaimed += reclaimed_now; - erts_smp_spin_unlock(&info_lck); - + esdp->gc_info.garbage_cols++; + esdp->gc_info.reclaimed += reclaimed_now; + FLAGS(p) &= ~F_FORCE_GC; #ifdef CHECK_FOR_HOLES @@ -504,10 +535,7 @@ erts_garbage_collect_hibernate(Process* p) /* * Preliminaries. */ - erts_smp_proc_lock(p, ERTS_PROC_LOCK_STATUS); - p->gcstatus = p->status; - p->status = P_GARBING; - erts_smp_proc_unlock(p, ERTS_PROC_LOCK_STATUS); + erts_smp_atomic32_read_bor_nob(&p->state, ERTS_PSFLG_GC); ErtsGcQuickSanityCheck(p); ASSERT(p->mbuf_sz == 0); ASSERT(p->mbuf == 0); @@ -618,9 +646,7 @@ erts_garbage_collect_hibernate(Process* p) ErtsGcQuickSanityCheck(p); - erts_smp_proc_lock(p, ERTS_PROC_LOCK_STATUS); - p->status = p->gcstatus; - erts_smp_proc_unlock(p, ERTS_PROC_LOCK_STATUS); + erts_smp_atomic32_read_band_nob(&p->state, ~ERTS_PSFLG_GC); } @@ -644,10 +670,7 @@ erts_garbage_collect_literals(Process* p, Eterm* literals, /* * Set GC state. */ - erts_smp_proc_lock(p, ERTS_PROC_LOCK_STATUS); - p->gcstatus = p->status; - p->status = P_GARBING; - erts_smp_proc_unlock(p, ERTS_PROC_LOCK_STATUS); + erts_smp_atomic32_read_bor_nob(&p->state, ERTS_PSFLG_GC); /* * We assume that the caller has already done a major collection @@ -789,9 +812,7 @@ erts_garbage_collect_literals(Process* p, Eterm* literals, /* * Restore status. */ - erts_smp_proc_lock(p, ERTS_PROC_LOCK_STATUS); - p->status = p->gcstatus; - erts_smp_proc_unlock(p, ERTS_PROC_LOCK_STATUS); + erts_smp_atomic32_read_band_nob(&p->state, ~ERTS_PSFLG_GC); } static int @@ -883,14 +904,12 @@ minor_collection(Process* p, int need, Eterm* objv, int nobj, Uint *recl) } } - if (wanted < MIN_HEAP_SIZE(p)) { - wanted = MIN_HEAP_SIZE(p); - } else { - wanted = next_heap_size(p, wanted, 0); - } + wanted = wanted < MIN_HEAP_SIZE(p) ? MIN_HEAP_SIZE(p) + : next_heap_size(p, wanted, 0); if (wanted < HEAP_SIZE(p)) { shrink_new_heap(p, wanted, objv, nobj); } + ASSERT(HEAP_SIZE(p) == next_heap_size(p, HEAP_SIZE(p), 0)); return 1; /* We are done. */ } @@ -1449,11 +1468,10 @@ adjust_after_fullsweep(Process *p, Uint size_before, int need, Eterm *objv, int I think this is better as fullsweep is used mainly on small memory systems, but I could be wrong... */ wanted = 2 * need_after; - if (wanted < p->min_heap_size) { - sz = p->min_heap_size; - } else { - sz = next_heap_size(p, wanted, 0); - } + + sz = wanted < p->min_heap_size ? p->min_heap_size + : next_heap_size(p, wanted, 0); + if (sz < HEAP_SIZE(p)) { shrink_new_heap(p, sz, objv, nobj); } @@ -1946,6 +1964,17 @@ setup_rootset(Process *p, Eterm *objv, int nobj, Rootset *rootset) ++n; } + /* + * A trapping BIF can add to rootset by setting the extra_root + * in the process_structure. + */ + if (p->extra_root != NULL) { + roots[n].v = p->extra_root->objv; + roots[n].sz = p->extra_root->sz; + ++n; + } + + ASSERT((is_nil(p->seq_trace_token) || is_tuple(follow_moved(p->seq_trace_token)) || is_atom(p->seq_trace_token))); @@ -1961,9 +1990,9 @@ setup_rootset(Process *p, Eterm *objv, int nobj, Rootset *rootset) n++; } #endif - ASSERT(is_nil(p->tracer_proc) || - is_internal_pid(p->tracer_proc) || - is_internal_port(p->tracer_proc)); + ASSERT(is_nil(ERTS_TRACER_PROC(p)) || + is_internal_pid(ERTS_TRACER_PROC(p)) || + is_internal_port(ERTS_TRACER_PROC(p))); ASSERT(is_pid(follow_moved(p->group_leader))); if (is_not_immed(p->group_leader)) { @@ -2523,6 +2552,12 @@ offset_one_rootset(Process *p, Sint offs, char* area, Uint area_size, p->dictionary->used, offs, area, area_size); } + if (p->extra_root != NULL) { + offset_heap_ptr(p->extra_root->objv, + p->extra_root->sz, + offs, area, area_size); + } + offset_heap_ptr(&p->fvalue, 1, offs, area, area_size); offset_heap_ptr(&p->ftrace, 1, offs, area, area_size); offset_heap_ptr(&p->seq_trace_token, 1, offs, area, area_size); @@ -2546,6 +2581,110 @@ offset_rootset(Process *p, Sint offs, char* area, Uint area_size, offset_one_rootset(p, offs, area, area_size, objv, nobj); } +static void +init_gc_info(ErtsGCInfo *gcip) +{ + gcip->reclaimed = 0; + gcip->garbage_cols = 0; +} + +static void +reply_gc_info(void *vgcirp) +{ + Uint64 reclaimed = 0, garbage_cols = 0; + ErtsSchedulerData *esdp = erts_get_scheduler_data(); + ErtsGCInfoReq *gcirp = (ErtsGCInfoReq *) vgcirp; + ErtsProcLocks rp_locks = (gcirp->req_sched == esdp->no + ? ERTS_PROC_LOCK_MAIN + : 0); + Process *rp = gcirp->proc; + Eterm ref_copy = NIL, msg; + Eterm *hp = NULL; + Eterm **hpp; + Uint sz, *szp; + ErlOffHeap *ohp = NULL; + ErlHeapFragment *bp = NULL; + + ASSERT(esdp); + + reclaimed = esdp->gc_info.reclaimed; + garbage_cols = esdp->gc_info.garbage_cols; + + sz = 0; + hpp = NULL; + szp = &sz; + + while (1) { + if (hpp) + ref_copy = STORE_NC(hpp, ohp, gcirp->ref); + else + *szp += REF_THING_SIZE; + + msg = erts_bld_tuple(hpp, szp, 3, + make_small(esdp->no), + erts_bld_uint64(hpp, szp, garbage_cols), + erts_bld_uint64(hpp, szp, reclaimed)); + + msg = erts_bld_tuple(hpp, szp, 2, ref_copy, msg); + if (hpp) + break; + + hp = erts_alloc_message_heap(sz, &bp, &ohp, rp, &rp_locks); + szp = NULL; + hpp = &hp; + } + + erts_queue_message(rp, &rp_locks, bp, msg, NIL +#ifdef USE_VM_PROBES + , NIL +#endif + ); + + if (gcirp->req_sched == esdp->no) + rp_locks &= ~ERTS_PROC_LOCK_MAIN; + + if (rp_locks) + erts_smp_proc_unlock(rp, rp_locks); + + erts_smp_proc_dec_refc(rp); + + if (erts_smp_atomic32_dec_read_nob(&gcirp->refc) == 0) + gcireq_free(vgcirp); +} + +Eterm +erts_gc_info_request(Process *c_p) +{ + ErtsSchedulerData *esdp = ERTS_PROC_GET_SCHDATA(c_p); + Eterm ref; + ErtsGCInfoReq *gcirp; + Eterm *hp; + + gcirp = gcireq_alloc(); + ref = erts_make_ref(c_p); + hp = &gcirp->ref_heap[0]; + + gcirp->proc = c_p; + gcirp->ref = STORE_NC(&hp, NULL, ref); + gcirp->req_sched = esdp->no; + erts_smp_atomic32_init_nob(&gcirp->refc, + (erts_aint32_t) erts_no_schedulers); + + erts_smp_proc_add_refc(c_p, (Sint32) erts_no_schedulers); + +#ifdef ERTS_SMP + if (erts_no_schedulers > 1) + erts_schedule_multi_misc_aux_work(1, + erts_no_schedulers, + reply_gc_info, + (void *) gcirp); +#endif + + reply_gc_info((void *) gcirp); + + return ref; +} + #if defined(DEBUG) || defined(ERTS_OFFHEAP_DEBUG) static int |